opencode-roundtable

OpenCode plugin that orchestrates multi-agent round-robin debates. Agents with different personalities debate a topic turn by turn, sharing context while keeping their own system prompts and tools.

A built-in observer automatically consolidates every debate into an executive summary.

Installation

npm (recommended)

{
  "plugin": ["@rrr2010/opencode-roundtable@latest"]
}

OpenCode auto-installs npm plugins on startup. No manual copy needed.

Local (dev)

git clone https://github.com/RrR2010/opencode-roundtable
cd roundtable
bun run build
npm link

Then add "@rrr2010/opencode-roundtable" to your opencode.json plugin array.

Features

• Round-robin debate — agents speak in sequence, each seeing the full discussion history
• Shared context — tool outputs and discoveries are visible to all participants
• Built-in observer — automatically consolidates the debate into an executive summary (overridable with a specific agent)
• Isolated session — the debate runs in a child session (like `task` tool), keeping the main session clean
• File persistence — state stored on disk, survives restarts
• Extend mode — continue a concluded roundtable with more rounds or a new topic
• Agent discovery — available_agents tool helps the orchestrator know which agents exist
• Active roundtables — active_roundtables tool lists current debates with status
• TUI command — /roundtables slash command lists sessions with clickable navigation
• Auto-navigate — optional auto-navigation between sessions on create/conclude
• Parallel roundtables — multiple independent debates can run simultaneously
• Loop detection — Jaccard bigram similarity detects agent impasses early

Tool API

roundtable()

roundtable({
  agents?: string[],    // Names in speaking order (min 2). Required for new debates.
  prompt: string,       // Topic or challenge. For multi-round, include per-round instructions.
  rounds?: number,      // Complete rounds (default: 1, max: 50)
  observer?: string,    // Agent for final consolidation (default: built-in)
  sessionID?: string,   // ses_xxxx — pass to extend a concluded debate
  title?: string,       // Custom title (default: auto-generated from prompt)
  observerPrompt?: string, // Override the observer consolidation prompt (e.g., "Save a detailed report to report.md")
})

Returns: { sessionID: string, summary: string } after the debate concludes.

Side effect: injects system-level prompts into each agent's context during the debate (role-setting, topic, turn routing, and lifecycle signals).

available_agents()

available_agents()

Returns: "Available agents: planner, developer, reviewer, ..." — a formatted string of agent names.

active_roundtables()

active_roundtables()

Returns: session IDs for agents to use programmatically (NOT clickable links). Only the /roundtables TUI command has clickable session navigation. Example:

Active roundtables:
- #ses_xxx · planner→developer→reviewer (R1/2) · debating
- #ses_yyy · planner→developer (R2/2) · consolidating

Usage Examples

Basic — one round, built-in observer

roundtable({
  agents: ["planner", "developer"],
  prompt: "What architecture should we use?",
})

Multi-round with per-round instructions

roundtable({
  agents: ["planner", "developer"],
  prompt: "Round 1: list pros. Round 2: list cons. Round 3: propose an implementation plan.",
  rounds: 3,
})

Explicit observer

roundtable({
  agents: ["planner", "developer"],
  prompt: "Should we migrate to microservices?",
  rounds: 2,
  observer: "reviewer",
})

Extend a concluded debate (preferred)

roundtable({
  sessionID: "ses_abc123",
  rounds: 2,
  prompt: "Dive deeper into operational costs",
})

Prefer extend over starting a new roundtable — it reuses accumulated context, saving exploration tokens. The session ID is shown in the S1 noReply message when the roundtable starts — check S1 context to find it.

Explicit per-agent per-round instructions

When you need each agent to focus on a specific aspect, reference them by name in the prompt:

roundtable({
  agents: ["planner", "developer", "reviewer"],
  rounds: 2,
  prompt: "Round 1 - planner: propose architecture. Round 2 - developer: implement plan. Round 2 - reviewer: review code.",
})

Discover agents first

const agents = available_agents()
// agents => "Available agents: planner, developer, reviewer, builder, architect"
roundtable({
  agents: ["planner", "developer"],
  prompt: "...",
})

List active roundtables

const active = active_roundtables()
// active => "Active roundtables:\n- #ses_xxx · planner→developer (R1/2) · debating"

Configuration

Place ~/.config/opencode/roundtable.json to override defaults:

{
  "$schema": "https://raw.githubusercontent.com/RrR2010/opencode-roundtable/refs/heads/master/docs/roundtable.schema.json",
  "defaultTimeoutMs": 300000,
  "loopSimilarityThreshold": 0.85,
  "toolOutputPreviewMax": 500,
  "maxRounds": 10,
  "defaultObserverPrompt": "You are an impartial roundtable observer...",
  "navigation": "link"
}

The JSON schema is available at docs/roundtable.schema.json.

If the file doesn't exist, it's created automatically with defaults.

Navigation modes

| Value | Behavior | |-------|----------| | "link" (default) | No auto-navigation. Relies on native #ses_xxx link rendering | | "auto" | Auto-navigates S1→S2 on create, S2→S1 on conclude | | "none" | No automatic navigation |

TUI Features

The plugin registers a /roundtables slash command that opens a dialog listing all active roundtables with clickable session navigation.

Tips

Multi-round strategy

For complex topics, structure rounds progressively:

1. Round 1: Explore — each agent lists their perspective
2. Round 2: Challenge — agents critique each other's positions
3. Round 3+: Converge — propose concrete solutions

Include all round instructions in the prompt parameter — all agents see the full agenda.

Nested roundtables

Recursive roundtables (an agent inside S2 calling roundtable() again) are prevented by a runtime guard. The plugin detects if the current session is already a roundtable and rejects the call with an error message. This ensures debates stay linear and predictable.

Extend mode

• Use the session ID from the S1 noReply message when the roundtable starts
• Original topic is preserved; new prompt becomes the continuation
• Agents must be the same as the original debate (validated)
• Continue an extend for iterative refinement

Documentation

• docs/SPEC.md — Full technical specification

Requirements

• OpenCode (latest version)
• No external dependencies

License

MIT